What you are actually paying for.
A service catalog is rarely evaluated as a line item, but it should be. Two to three engineers spend real, recurring time keeping a hand-maintained catalog running: writing YAML, chasing owners, reconciling entries that were already wrong before the pull request merged. That cost repeats every quarter, whether or not anyone notices.
The decision is not whether to have a catalog. Every team already has one. The decision is whether it should keep being maintained by hand.
The reader changed: from humans to agents.
A stale catalog entry used to cost an engineer twenty minutes of confusion, then they noticed and corrected course. Coding, deployment, and incident-response agents now read the same entry and act on it, with no built-in moment of doubt. The maintenance problem became a reliability problem.
A copilot that reads a stale catalog gives bad advice. An agent that acts on one causes an incident. As your team ships production agents, catalog accuracy stops being a convenience and becomes a reliability decision your roadmap depends on.
The maintenance tax, quantified.
Most teams calculate the cost of a catalog in engineering hours and stop there. The fuller model has three lines: headcount spent on curation, incident minutes lost to topology discovered mid-outage, and the agent-risk exposure of every automated system that reads the catalog and trusts it.
Incident minutes are the easiest to underweight and the most expensive to ignore, because they land during your worst moments, not during a calm sprint retro.
Observed, not declared.
Every catalog tool on the market, Backstage, Cortex, Compass, and Roadie, shares the same architecture: a human declares the truth, usually in YAML, and the tool trusts that declaration until someone updates it. NOFire AI catalogs from observation instead: an in-cluster agent, your cloud provider integrations, your code, and your telemetry are synthesized into a single Production Context Graph, so the catalog reflects what is actually running.
Today: declared. Backstage, Cortex, and Compass require humans to keep it accurate, and it decays the moment they stop. NOFire: observed. Synthesized continuously from the cluster, the cloud, the code, and production telemetry, current without anyone touching it, and with no YAML.
What you get, and how you trust it.
Every fact in the catalog carries a provenance label, runtime, synthesized, or intent, each with a confidence score, so humans and agents know what to trust before they act on it. The catalog is seeded from your GitHub repos, workflows, and release tags, so it is never empty before the first metric arrives. Past incident resolutions are captured and surfaced automatically, so the catalog gets smarter after every incident.
Migration and a 90-day rollout.
Staying on Backstage has a running cost. Leaving it does not have to be a project. A practical rollout: connect the in-cluster agent, your cloud provider, and your repositories in week one, let those signals synthesize the catalog alongside your existing portal, validate the observed graph against the topology your team already trusts, then retire the manual entries. It runs alongside Backstage during the transition, so there is no cutover weekend and nothing to migrate.
The vendor landscape, by cost.
- Backstage: free to license, expensive to run. Six to twelve months to a usable portal, and one to two dedicated platform engineers once you cross roughly 200 engineers.
- Cortex: per-seat pricing plus professional services, plus an estimated 0.25 FTE spent on ongoing catalog hygiene to prevent descriptor rot.
- Roadie: a $1,200 per month, 50-seat floor before a single developer opens the portal, with RBAC, the REST API, and SLA guarantees gated behind a 100-plus seat tier.
- NOFire AI: no seat minimum, no per-engineer hygiene tax, and no infrastructure to run.
The Compass deadline, and what it means for your roadmap.
Atlassian ended Compass sales on May 13, 2026. Support ends December 31, 2027. Teams that already migrated once, from Opsgenie to Compass, are now being asked to migrate again, to a product with no proven feature parity.
Migrating to another hand-maintained catalog just resets the clock on the same decay. NOFire AI has nothing to migrate: no YAML to export, no entries to re-create, and a catalog that maintains itself from the day it connects.
For the full architecture, the provenance model, and the chapter-by-chapter breakdown of each alternative, see the Service Catalog Guide. To see the approach applied to your own stack, visit the self-maintaining service catalog.